PECA v3: Evidence-Gated Anticipatory Execution: Separating Requirement Compliance, Predictive Utility, and Commit Authorization in AI Agents

@misc{rm3692026peca,
  author       = {{RM369}},
  title        = {{PECA v3: Evidence-Gated Anticipatory Execution: Separating Requirement Compliance, Predictive Utility, and Commit Authorization in AI Agents}},
  year         = {2026},
  month        = oct,
  date         = {2026-10-26},
  publisher    = {Life of Research},
  howpublished = {Preprint},
  url          = {https://lifeofresearch.com/preprints/140},
  abstract     = {Anticipation can help an AI agent prepare useful work before a requirement is fully specified, but a predictor can also create the very obligations the agent then tries to satisfy. PECA v3 refactors Prefix-Conditioned Execution and Cumulative Audit around a strict separation among confirmed requirements, provisional predictions, and the mutable artifact with its provenance. The controller preserves the state triple (R\_h,R\_p,S), but makes the distinction operational: compliance is judged only against active, evidence-supported requirements in R\_h; predictions in R\_p are evaluated separately for utility, calibration, and waste; and external or consequential changes pass through action-specific authorization and commit gates. Each action carries a contract covering evidence, preconditions, expected effects, verification, side effects, reversibility, and commit criteria. The design favors preparation useful across several plausible futures, permits isolated branch-specific speculation only when its expected value is positive, and combines material-cycle checkpoints with event-triggered reviews. The nested 3-6-9 cadence is retained as an ablatable scheduling hypothesis, not a source of correctness. This paper defines the architecture and a controlled evaluation plan. It reports no live-model performance results and makes no claim that the proposed controller is superior until matched, independently evaluated experiments establish that fact.},
  keywords     = {AI agents, instruction following, speculative execution, runtime verification, auditability, action contracts, reversibility, reward hacking},
  language     = {en},
  copyright    = {Creative Commons Attribution 4.0}
}